def make_generator_inputs(self, data):
source = preproc.mr2source_inputs(data)
tokens = [self.source_vocab.start_token] + source \
+ [self.source_vocab.stop_token]
inputs = Variable(
torch.LongTensor([[self.source_vocab[t] for t in tokens]]).t(),
lengths=torch.LongTensor([len(tokens)]),
length_dim=0, batch_dim=1,
pad_value=self.source_vocab.pad_index)
if self._gpu > -1:
inputs = inputs.cuda(self._gpu)
return {"source_inputs": inputs}
def run(self, env, verbose=False):
output_path = env["proj_dir"] / "output" / self.filename
output_path.parent.mkdir(exist_ok=True, parents=True)
if self.checkpoint is None:
ckpt = self._get_default_checkpoint(env)
else:
ckpt = self.checkpoint
if ckpt is None:
raise RuntimeError("No checkpoints found!")
ckpt_path = env["checkpoints"][ckpt]["path"]
if verbose:
print("Loading model from {}".format(ckpt_path))
model = plum.load(ckpt_path).eval()
if env["gpu"] > -1:
model.cuda(env["gpu"])
self._gpu = env["gpu"]
samples = self.make_samples()
with open(output_path, "w") as out_fp:
for i, mr in enumerate(samples, 1):
print("{}/{}".format(i, len(samples)),
end="\r" if i < len(samples) else "\n",
flush=True)
gen_input = self.make_generator_inputs(mr)
source = preproc.mr2source_inputs(mr)
tokens = self._get_outputs(model, gen_input)
data = json.dumps({
"source": source,
"mr": mr,
"text": " ".join(tokens),
})
print(data, file=out_fp, flush=True)